Deep Convolutional Neural Network for Anomaly Detection in Power System Networks

Nowadays, the increasing complexity of power system networks has made them valuable to various types of anomalies. Also, the self-organizing neural networks learn patterns to recognize in data without prior knowledge of the underlying system as a solution for anomaly detection in power system networks. However, the existing One-Class Support Vector Machine (OC-SVM) faced challenges in detecting anomalies due to the high-dimensional and non-linear nature of data. Hence, this research proposes Deep Convolutional Neural Network (DCNN) to detect anomalies in traffic real-time data. Initially, the input data is collected from KDDcup99 dataset which consists of various Daniel of Service (DoS) attack data points. Then, the data is preprocessed with Regression Imputation (RI) to handle the missing and duplicate data. After that, Fisher Discriminant Analysis (FDA) is used to extract the features by finding the optimal linear combination of features that maximizes the separation between normal and attack classes. Finally, the proposed DCNN is introduced to detect the anomalies by learning hierarchical representations of temporal and spatial patterns. From the results, the proposed DCNN established better results when compared to Long Short-Term Memory with Auto Encoders (LSTM-AE) in terms of precision (96.72%) and recall (98.76%) respectively.